Speak at The Fifth Elephant 2026 Annual Conference

Description

Unlike coding or conversational question answering, support request resolution has no generic workflow that an AI agent can follow. Resolution paths vary across customers and domains (for example, using Okta vs. IdentityNow for access requests), involve constantly changing context, require high-risk actions (such as granting admin access or removing users from licenses), and must strictly adhere to customer-specific policies (for example, setting a ticket status to “Triage” when routing it to another team).

We present a Supervisor + Specialists agentic architecture designed specifically for this domain. The system maintains a shared ticket state, generates hierarchical execution plans using plan graphs, dynamically replans based on validation signals, and orchestrates multiple specialist skills through a scalable custom agent harness. Tool execution is verified using preconditions and postconditions to ensure that actions have the intended effect.

Outline

Why is support resolution uniquely hard?

• Support tickets are often underspecified.
• Context changes while the agent is executing.
• Mistakes can have significant security or availability consequences.
• Producing a good textual response alone is not sufficient.

Multi-Agent Architecture: Supervisor + Specialists

• A supervisor agent maintains a shared ticket state.
• Work is delegated to domain-specific specialist agents.
• Specialists communicate using well-defined inputs, outputs, and contextual information.

Agent Planning Strategies

• Hierarchical planning using plan graphs.
• Explicit re-planning triggers based on plan validation and context enrichment.

Tool and Skill Execution

• Every tool call is wrapped with typed schemas.
• Preconditions validate whether an action should be executed.
• Postcondition checks verify that the intended system state has actually changed.

Operating an Agentic System at Scale

• Handling more than 30 million support requests while managing token usage and infrastructure costs.
• Dealing with burst traffic and noisy-neighbor scenarios.
• Engineering reliable, large-scale production systems.

Customer Impact

• The agentic resolution system currently resolves approximately 40% of customer support requests.
• Human-AI collaboration through human-in-the-loop workflows and continuous learning pipelines.

Takeaways

• How agentic systems can create measurable business impact in enterprise environments.
• Engineering challenges involved in building reliable, high-quality agentic systems at scale.
• Difficulties in objectively measuring agent quality and correlating it with business outcomes.
• Strategies for continuously improving agent performance by incorporating human feedback.

Who Should Attend?

This talk is intended for a broad audience, including:

• AI researchers
• AI developers and engineers
• Software engineers building production AI systems
• Product managers
• Business and engineering leaders

It is particularly relevant for anyone interested in large-scale agentic applications and how AI agents can deliver meaningful customer impact.

{Add your bio - who you are; where you work}

{Add the link to draft slides - PDF/PPT - with comments access}

{Add the link to 2-min elevator pitch video}